Aimed at the complex work environment of Agricultural robots, this paper introduced a new robust feature point detection algorithm - SURF algorithm. It shows good robustness in environment where exist illumination, rotation, or scale change, and achieve sub-pixel precision level; On this basis, we used the nearest neighbor rule and BBF (best bin first) search algorithm to do many feature points matching experiments. The result shows that the method is more robust than Harris and has better real time than SIFT algorithm. It has some theoretical and practical value in robot vision localization, map building, intelligent navigation, etc.